🦾 GPT Astra for robot is impressive.. but slow! Can we use light-weight Qwen for real-time control, or use any general VLMs like Gemini, Claude? 🕹️ We are thrilled to introduce Show-Harness, a semantic action interface that brings the physical intelligence from a general VLM agent into robot-use. No more calibration. No extra VLA or SAM. 🧠 Show-Harness generalizes across tasks and scenes, using GPT, Gemini, Claude zero-shot without training, or just fine-tuning Qwen with 164 demos within 1 hour. Model and code👇 🌐 Project Page: showlab.github.io/Show-Harne… 📄 Paper: arxiv.org/abs/2609.10522 💻 Code: github.com/showlab/Show-Harn… 🤗 Model & Data: huggingface.co/collections/s…

After pivoting to embodied AI last year, this is our first time submitting to #CoRL 2026, and I’m excited to share that Show Lab has 3 out of 3 papers accepted! 🎉 Huge congratulations to the students behind these papers — all were 1st-year PhD or master students, completely new to the field, who spent tremendous efforts to pick up the know-how of robot learning 🤖 1. Supervise What Survives: Geometry-Guided VLA Adaptation from Synthetic Robot Videos How robot can learn from human demo video? Video gen model can edit human video as robot video, but how to use these generated human-edit-as-robot video for training robot? Instead of hoping to learn how to control, we propose to guide robot learning with the geometry in the human-edit-as-robot video. 2. Where Success Breaks: Failure-Boundary Learning for Robust Vision-Language-Action Models When training VLAs, we often focus on successful executions. But failures contain equally important information about where a robot is starting to go wrong. We introduce Failure-Boundary Learning, reframing robust VLA adaptation as the problem of discovering, localizing, and shaping the boundary between recoverable deviations and task failure. 3. From Video Frames to Robot Meta-Frames: A Unified Latent Interface for World Action Models World Action Models such as Cosmos Policy show great potential for jointly predicting actions and video. However, different robot signals — images, proprioception, actions, future states, and values — are typically represented as separate images, making inference computationally costly. We introduce MetaWAM, which packs these different modalities into a single unified meta-frame in the latent space. On RoboCasa, MetaWAM achieves 68.9% success, while reducing inference latency by 39%.

How robot can learn from human demo video? Video gen model can edit human video as robot video, but how to use these generated human-edit-as-robot video for training robot? Congrats @DanzerChan and the team on the #CoRL paper "Supervise What Survives" arxiv.org/abs/2606.24448 🎊 which proposes to guide robot learning with geometry in the human-edit-as-robot video.